Attach spectral decompositions to a list of dense kernels

Turns each dense kernel (list(Kernel = <n x n>, Type = "D"|"BD")) into a cached spectral decomposition stored on $decomp. Call once and reuse across any number of kernel_model fits: kernels depend only on genotypes and environments – never on the phenotype vector – so masking y for a cross-validation fold does NOT invalidate the decomposition.

Usage

decompose_kernels(
  K,
  ne = NULL,
  tol = 1e-10,
  scale_kernels = TRUE,
  keep_var = 1,
  verbose = TRUE
)

Arguments

K

list of kernels.

ne

per-environment observation counts; defaults to attr(K, "ne").

tol

eigenvalue cutoff; must match kernel_model(tol=).

scale_kernels

must match kernel_model(scale_kernels=).

keep_var

numeric in (0, 1]. Retain only the leading eigenvalues carrying this proportion of each kernel's trace. 1 (default) keeps everything. A named vector or list keyed by kernel name / regex applies per kernel, e.g. c(KGE_ = 0.95).

verbose

report rank, timing and route per kernel.

Value

K with $decomp on each element and a "kd_meta" attribute.

Details

Factored kernels from get_kernel arrive pre-decomposed and do not need this function.

See also

get_kernel, kernel_model, undecompose_kernels